Ultraviolet plume target detection method based on transformer

Abstract

The application discloses a UV plume target detection method based on a Transformer, and belongs to the technical field of image recognition, which comprises a first stage of pipeline filtering for fast feature recognition and judgment, and a second stage of constructing a Transformer model based on joint space-time features for further distinguishing the survival samples filtered in the first stage, so as to effectively distinguish the UV plume target from interference sources such as arc discharge in a low-altitude complex UV clutter environment, thereby reducing the false alarm rate and improving the detection efficiency. The application can quickly eliminate noise and most random interference sources, effectively reduce the false alarm rate, accurately depict the difference between the UV plume target and the interference source in dynamic evolution, thereby improving the distinguishing ability, achieving the compromise between low false alarm rate and real-time performance, and being suitable for engineering application in a low-altitude complex environment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image recognition, and particularly relates to a UV plume target detection method based on a Transformer. BACKGROUND

[0002] The UV plume target detection mainly utilizes the characteristics of the "solar blind" UV band (200nm-300nm). Due to the absorption of ozone, the solar radiation transmitted to the earth's surface in this band is extremely weak, and the background signal reflected back to the detector by the ground object is smaller. Therefore, when a specific threat target generates a large amount of plume in the active flight stage, it will be accompanied by a relatively strong solar blind UV radiation signal, and the signal strength is relatively high, and the signal-to-noise ratio is good, which is conducive to target detection.

[0003] However, in the low-altitude environment, human social activities have caused a complex UV clutter background. For example, electric arcs, welding and various arc discharge light sources in industrial production can produce strong UV radiation, which becomes the main interference source of the UV detection system, seriously affecting the accuracy of detection.

[0004] In the existing UV plume target detection method, pipeline filtering is a commonly used means. This method uses the continuity of target motion in space and the statistical radiation characteristics of UV plume to distinguish targets, effectively eliminating random noise and most irregular flickering interference sources. However, in the low-altitude complex environment, the radiation mode of part of the interference source in a short time (such as 1 second) is similar to that of the UV plume target, and the pipeline filtering is difficult to distinguish, thereby generating a high false alarm rate. SUMMARY

[0005] The purpose of the present application is to provide a UV plume target detection method based on a Transformer, which solves the technical problem of effectively distinguishing UV plume targets from arc discharge and other interference sources in a low-altitude complex UV clutter environment, thereby reducing the false alarm rate and improving the detection efficiency.

[0006] To achieve the above purpose, the present application adopts the following technical scheme:

[0007] The UV plume target detection method based on a Transformer includes the following steps:

[0008] Step 1: The data acquisition module obtains a continuous image sequence through the UV imaging device, and sends the image sequence to the pipeline filtering module. The pipeline filtering module performs first-stage pipeline filtering on the image sequence, performs primary detection on the suspected target according to the motion continuity and statistical radiation characteristics of the UV plume target, obtains a survival sample, constructs a survival sample set and outputs;

[0009] Step 2: The size regularization module calls the set of survival samples, calls the survival sample, sets a unified interception window according to the position information and size information of the survival sample in space-time, sets the size of the interception window as C, and performs size regularization processing on the survival sample by using the interception window;

[0010] Step 3: In the interception window, the space-time interception module intercepts the image region centered on the survival sample from the continuous 90 frames of images to form a 90-frame pixel matrix array, reduces the spatial size of the pixel matrix array by using the bilinear interpolation method, constructs an array containing 90 64x64 pixel matrices, and obtains the space-time data X of the survival sample;

[0011] Step 4: The feature extraction module calls the space-time data X and inputs it into the convolution layer to extract an 8-dimensional spatial feature vector representing the spatial gray scale features of the ultraviolet plume target;

[0012] Step 5: The Transformer encoding module performs class embedding on the extracted 8-dimensional feature vector, then superimposes the position coding, and inputs it into the Transformer encoder for deep space-time feature modeling;

[0013] Step 6: The classification and discrimination module takes the class embedding corresponding position from the features output by the Transformer encoder, passes through the fully connected layer, and obtains the final classification result. According to the classification result, it is judged whether the survival sample is an ultraviolet plume target or an interference.

[0014] Preferably, when step 1 is executed, the pipeline filtering module performs the following steps on the received continuous image sequence:

[0015] Step 1-1: Extract the motion trajectory of the candidate pixel points between the continuous image frames, compare the position changes of the target in the time dimension, and use the continuous motion characteristics of the ultraviolet plume target to eliminate isolated noise points in a single image frame;

[0016] Step 1-2: Extract the ultraviolet radiation intensity in the candidate region and count the radiation features to distinguish noise and suspected targets;

[0017] Step 1-3: According to the motion continuity characteristics and radiation features of the ultraviolet plume target, use the moving pipeline detection method to retain only the candidate targets that exist continuously in N frames;

[0018] Step 1-4: According to the result of step 1-3, generate a survival sample and record the feature information of the survival sample:

[0019] Detection time: detection frame index T;

[0020] Number of frames existing in the moving pipeline: number of existing frames N;

[0021] spatiotemporal position: center coordinates (x t , y t ) of the suspected object in the frame image, object size (w t , h t ), wherein w represents width and h represents height, and t is frame index;

[0022] frame index: t = T - N + 1,..., T - 1, T;

[0023] Step 1-5: generating a set of survival samples according to the result of step 1-4.

[0024] Preferably, when performing step 2, the following steps are specifically included:

[0025] Step 2-1: for each survival sample, taking 90 frames before its detection frame T, i.e. t = T - 89,..., T, collecting the width sequence {w t} and height sequence {h t} of the object;

[0026] If the survival sample has a frame number N < 90, the size of the target in the insufficient frames is set to 0;

[0027] Step 2-2: calculating the maximum value of the width sequence:

[0028] w max = max ({w t});

[0029] calculating the maximum value of the height sequence:

[0030] h max = max ({h t});

[0031] Step 2-3: determining the side length C of the intercepting window:

[0032] if w max ≤ 64 and h max ≤ 64, then set C = 64; otherwise, set C = max (w max , h max );

[0033] Step 2-4: generating a square intercepting window with a side length C for each survival sample according to the method of step 2-1 to step 2-3.

[0034] Preferably, when performing step 3, the following steps are specifically included:

[0035] Step 3-1: For each survival sample, according to the detection frame index T, extract the first 89 frames and the detection frame of the survival sample, that is, a total of 90 continuous images are extracted, and the specific frame index is:

[0036] t = T - 89,..., T - 1, T;

[0037] Step 3-2: In each frame image I t , a square region with a side length of C is intercepted as an interception window with the center coordinates (x t , y t ) of the survival sample as the center:

[0038] S t = I t [(x t - C / 2): (x t + C / 2), (y t - C / 2): (y t + C / 2)];

[0039] Wherein, St is the spatial data of frame image It;

[0040] Step 3-3: If the interception window exceeds the image boundary, the exceeding part is filled with 0 value, and if the survival sample has less than 90 frames, the missing frame data is filled with a full 0 matrix;

[0041] Step 3-4: Combine the spatial data of continuous frames to obtain an array S containing 90 Cx C pixel matrices:

[0042] S = [S T−89 ,..., S T−1 , S T ];

[0043] Step 3-5: According to the size of the interception window side length C, judge whether to perform spatial reduction processing:

[0044] If C = 64, take the space-time data X = S;

[0045] If C > 64, perform bilinear interpolation scaling on each frame image matrix S t to adjust the size to 64x64, and obtain new space-time data X:

[0046] X = [↓S T−89 ,..., ↓S T−1 , ↓S T ];

[0047] Wherein, ↓ represents the downsampling operation of bilinear interpolation;

[0048] Step 3-6: Output the space-time data X.

[0049] Preferably, step 4 includes the following steps:

[0050] Step 4-1: Retrieve spatiotemporal data X;

[0051] Step 4-2: Stitch the spatiotemporal data X horizontally to obtain a large image X. img :

[0052] X img =hstack(X);

[0053] Where hstack is the horizontal concatenation operation;

[0054] Step 4-3: Place X img Remodeled into tensor X tensor :

[0055] X tensor =reshape(X img ,(1,1,64,5760));

[0056] Where, reshape is the reshaping operation, and (1,1,64,5760) are the parameters for reshaping;

[0057] Step 4-4: For tensor X tensor Perform convolution to extract 8-dimensional feature X conv :

[0058] X conv =Conv2d(X tensor ,out_channels=8,kernel_size=64,stride=64);

[0059] Conv2d is a 2D convolution with a kernel size of 64, an output channel count of 8, and a stride of 64.

[0060] Steps 4-5: Convert the 8-dimensional feature X conv Rearrange as The tensor X = 90 × 8 seq :

[0061] X seq =rearrange(X conv ,′bchw→b(hw)c′);

[0062] Where rearrange is the rearrangement operation, 'bchw→b(hw)c' is the parameter of the rearrangement operation, and tensor X seq It contains spatiotemporal features of 90 frames, each frame is represented by an 8-dimensional feature vector.

[0063] Preferably, step 5 includes the following steps:

[0064] Step 5-1: In tensor X seq Concatenate a category embedding to obtain tensor X cls :

[0065] X cls =Concat(CLS,X seq );

[0066] Where Concat is the concatenation operation and CLS is the learnable category embedding;

[0067] Step 5-2: Superimpose position information to obtain tensor X pos :

[0068] X pos =X cls +POS;

[0069] Where POS stands for location code;

[0070] Step 5-3: Place X pos The input is fed into a Transformer encoder for deep spatiotemporal feature modeling, yielding the output X of the Transformer encoder. trans :

[0071] X trans =Transformer(X pos );

[0072] Here, Transformer is the Transformer encoder.

[0073] Preferably, step 6 includes the following steps:

[0074] Step 6-1: Retrieve the Transformer encoder output X trans ;

[0075] Step 6-2: Output F at the first position corresponding to the category embedding:

[0076] F=X trans [:,0];

[0077] Among them, X trans [:,0] indicates the output of the first position;

[0078] The data is fed into a fully connected layer for classification, and the classification result y is output:

[0079] y = argmax(FC1(F));

[0080] wherein FC1 is a single-layer fully connected layer, and argmax is an index of taking maximum number;

[0081] y=1 represents the ultraviolet plume target, and y=0 represents interference.

[0082] The ultraviolet plume target detection method based on the Transformer solves the technical problem of effectively distinguishing the ultraviolet plume target from interference sources such as arc discharge in a low-altitude complex ultraviolet clutter environment, thereby reducing the false alarm rate and improving the detection efficiency. The two-stage detection architecture is adopted in the present application. In the first stage, pipeline filtering is used to realize the rapid elimination of noise and most random interference sources by using the target motion continuity and radiation statistical characteristics. In the second stage, a Transformer model based on joint modeling of space-time features is constructed for the suspected samples that survive the first stage, which is specially designed to identify low-altitude complex interference sources and effectively reduce the false alarm rate. Instead of directly modeling the entire video, the present application uses the precise space-time positioning provided by the pipeline filtering to only intercept the corresponding surviving sample segments as the input of the Transformer, significantly reducing the input data volume. The present application uses a single-layer convolution to extract an 8-dimensional spatial gray feature vector to represent each frame of data based on the precise positioning characteristics of the target, which not only preserves key morphological information but also significantly saves computing resources. The Transformer model combines spatial gray features and time series transformation patterns to accurately depict the differences in dynamic evolution between ultraviolet plume targets and interference sources, thereby improving the discrimination ability. Through the two-stage design of “rough screening first and then precise judgment”, the low false alarm rate and real-time performance are balanced, which is suitable for engineering applications in low-altitude complex environments. BRIEF DESCRIPTION OF DRAWINGS

[0083] Figure 1 is the main flowchart of the present application;

[0084] Figure 2 is the flowchart of step 1 of the present application;

[0085] Figure 3 is the flowchart of step 2 of the present application;

[0086] Figure 4 is the flowchart of step 3 of the present application;

[0087] Figure 5 is the flowchart of step 4 of the present application;

[0088] Figure 6 is the flowchart of steps 5 and 6 of the present application. DETAILED DESCRIPTION

[0089] from Figures 1-6The illustrated Transformer-based ultraviolet plume target detection method includes the following steps:

[0090] Step 1: The data acquisition module obtains a continuous image sequence through the ultraviolet imaging device, and sends the image sequence to the pipeline filtering module. The pipeline filtering module performs first-stage pipeline filtering on the image sequence, preliminarily detects suspected targets according to the motion continuity and statistical radiation characteristics of the ultraviolet plume target, obtains survival samples, constructs a survival sample set, and outputs the survival sample set.

[0091] When performing step 1, the pipeline filtering module performs the following steps on the continuous image sequence after receiving the continuous image sequence:

[0092] Step 1-1: Extract the motion trajectory of the candidate pixel points between the continuous image frames, compare the position changes of the target in the time dimension, and remove isolated noise points in a single image frame by using the continuous motion characteristics of the ultraviolet plume target.

[0093] Step 1-2: Extract the ultraviolet radiation intensity in the candidate region, and statistically analyze the radiation characteristics to distinguish noise and suspected targets.

[0094] Step 1-3: According to the motion continuity characteristics and radiation characteristics of the ultraviolet plume target, a moving pipeline detection method is used to retain only the candidate targets that exist continuously in N frames.

[0095] Step 1-4: According to the results of step 1-3, generate survival samples and record the feature information of the survival samples:

[0096] Detection time: detection frame index T;

[0097] Number of frames existing in the moving pipeline: number of existing frames N;

[0098] Space-time position: Center coordinates (x t , y t ) of the suspected target in the frame image, target size (w t , h t ), where w represents width, h represents height, and t is the frame index;

[0099] Frame index: t = T-N+1,..., T-1, T;

[0100] Step 1-5: According to the results of step 1-4, generate a survival sample set.

[0101] In the first stage, pipeline filtering is used to perform the first stage coarse detection to exclude noise and most of the interference. The pipeline filtering utilizes the motion continuity of the ultraviolet plume target and the statistical radiation characteristics to detect the suspected target, i.e., the survival sample in the first stage. Each survival sample has a detection time, which is represented by the index T of the detection frame. If the survival sample exists in the moving pipeline for N frames, the spatiotemporal position of the survival sample is represented by the center coordinates (x t , y t ) of the suspected target in the N frames of images and the target size (w t , h t ), where the frame index t = T-N+1,..., T-1, T.

[0102] Step 2: The size regularization module calls the survival sample set, calls the survival sample therein, sets a unified clipping window according to the spatiotemporal position information and size information of the survival sample, sets the size of the clipping window as C, and performs size regularization processing on the survival sample by using the clipping window.

[0103] When step 2 is performed, the following steps are specifically included:

[0104] Step 2-1: For each survival sample, 90 frames before the detection frame T, i.e., t = T-89,..., T, are taken to collect the width sequence {w t} and the height sequence {h t} of the target.

[0105] If the survival sample exists for less than 90 frames, the target size in the insufficient frames is set as 0.

[0106] Step 2-2: The maximum value of the width sequence is calculated:

[0107] w max =max({w t});

[0108] The maximum value of the height sequence is calculated:

[0109] h max =max({h t});

[0110] Step 2-3: The side length C of the clipping window is determined:

[0111] If w max ≤64 and h max ≤64, C is set as 64; otherwise, C is set as max(w max , h max ).

[0112] Step 2-4: Following the methods in Steps 2-1 to 2-3, a square cut-out window with side length C is generated for each surviving sample.

[0113] For each survival sample, this step sets a square truncating window with side length C based on its spatiotemporal location, i.e., performs truncating size regularization. The side length C is calculated as follows. For this survival sample, 90 frames before the detected frame T are taken to obtain the size sequence {(w t h t )},

[0114] t = T-89,...,T-1,T; If the number of frames N in which the survival sample exists is less than 90, the size of the survival sample in the insufficient frames is set to 0. Calculate the {w} of the width sequence. t The maximum value of} is w max , and the height sequence {h t The maximum value of} is h max If w max and h max If none of them exceed 64, then the cut-off size C is set to... Otherwise, set the cutoff size C to w. max and h max The larger values.

[0115] Step 3: Within the capture window, the spatiotemporal capture module captures the image region centered on the survival sample from 90 consecutive frames of images, forming a 90-frame pixel matrix array. The spatial size of the pixel matrix array is reduced using bilinear interpolation to construct an array containing 90 64×64 pixel matrices, thus obtaining the spatiotemporal data X of the survival sample.

[0116] Step 3 includes the following steps:

[0117] Step 3-1: For each survival sample, extract the first 89 frames and the detected frames based on the detected frame index T, resulting in a total of 90 consecutive images. The specific frame indices are as follows:

[0118] t = T−89,…,T−1,T;

[0119] Step 3-2: In each frame image I t Above, using the center coordinates (x) of the survival samples t y t Using a point as the center, extract a square region with side length C as the extraction window:

[0120] S t =I t [(x t −C / 2):(x t +C / 2),(yt −C / 2):(y t +C / 2)];

[0121] Where St is the spatial data of frame image It;

[0122] Step 3-3: If the cropping window exceeds the image boundary, the excess part is filled with 0 values. If the number of survival samples is less than 90 frames, the missing frame data is filled with an all-zero matrix.

[0123] Steps 3-4: Combine the spatial data of consecutive frames to obtain an array S containing 90 C×C pixel matrices:

[0124] S=[S T−89 ,...,S T−1 ,S T ];

[0125] Steps 3-5: Based on the size of the side length C of the captured window, determine whether to perform space reduction processing:

[0126] If C=64, then take the spatiotemporal data X=S;

[0127] If C > 64, then for each frame of the image, the matrix S t Perform bilinear interpolation scaling to adjust the size to 64×64, resulting in new spatiotemporal data X:

[0128] X=[↓S T−89 ,...,↓S T−1 ,↓S T ];

[0129] Where ↓ represents the downsampling operation of bilinear interpolation;

[0130] Steps 3-6: Output spatiotemporal data X.

[0131] For each surviving sample, this step extracts spatiotemporal data from consecutive image frames based on its spatiotemporal location and the cropping size C, forming a dataset containing 90 images. An array of 4×64 pixel matrices. The specific extraction method is as follows. For this survival sample, 90 frames before the detected frame T are extracted, using the survival sample coordinates (x... t y t Centered on frame I, a square cropping window with side length C is extracted, containing the image data. That is, for the t-th frame image I... t Given t = T-89, ..., T-1, T, take a C×C window region of the image to obtain the spatial data S of that frame. t .

[0132] Step 4: The feature extraction module calls the space-time data X and inputs it into a convolution layer to extract an 8-dimensional space feature vector representing the spatial gray scale features of the ultraviolet plume target.

[0133] In step 4, the following steps are included:

[0134] Step 4-1: Call the space-time data X;

[0135] Step 4-2: Concatenate the space-time data X in the horizontal direction to obtain a large image X img :

[0136] X img =hstack(X);

[0137] Where hstack is a horizontal concatenation operation;

[0138] Step 4-3: Reshape X img to a tensor X tensor :

[0139] X tensor =reshape(X img ,(1,1,64,5760));

[0140] Where reshape is a reshape operation, and (1,1,64,5760) is the reshape parameter;

[0141] Since the survival samples can be accurately positioned in space, a convolution layer is used in this embodiment to extract an 8-dimensional feature vector to represent the spatial gray scale features of the ultraviolet plume target. The specific extraction method is as follows. An array X containing 90 64x64 pixel matrices is concatenated in the horizontal direction to form a large 64x5760 image, and then reshaped into a 1x1x64x5760 tensor.

[0142] Step 4-4: Convolve the tensor X tensor to extract an 8-dimensional feature X conv :

[0143] X conv =Conv2d(X tensor ,out_channels=8,kernel_size=64,stride=64);

[0144] Where Conv2d is a 2-dimensional convolution, the convolution kernel kernel_size is 64, the output channel number out_channels is 8, and the step stride is 64;

[0145] Step 4-5: Rearrange the 8-dimensional feature X conv to​ Tensor X of 90x8 seq :

[0146] X seq = rearrange(X conv , 'bchw->b(hw)c');

[0147] wherein rearrange is a rearrangement operation, 'bchw->b(hw)c' is a parameter of the rearrangement operation, and tensor X seq contains 90 frames of spatio-temporal features, each frame being represented by an 8-dimensional feature vector.

[0148] Step 5: The Transformer encoding module performs class embedding on the extracted 8-dimensional feature vector, then superimposes position encoding, and inputs the result into a Transformer encoder for deep spatio-temporal feature modeling.

[0149] When performing step 5, the following steps are specifically included:

[0150] Step 5-1: Concatenate a class embedding in front of tensor X seq to obtain tensor X cls :

[0151] X cls = Concat(CLS, X seq );

[0152] wherein Concat is a concatenation operation, and CLS is a learnable class embedding.

[0153] Step 5-2: Superimpose position information to obtain tensor X pos :

[0154] X pos = X cls + POS;

[0155] wherein POS is position encoding.

[0156] Step 5-3: Input X pos into a Transformer encoder for deep spatio-temporal feature modeling to obtain the output result X trans of the Transformer encoder:

[0157] X trans = Transformer(X pos );

[0158] wherein Transformer is a Transformer encoder.

[0159] Step 6: The classification discrimination module takes the category embedding corresponding position from the features output by the Transformer encoder, and after passing through the fully connected layer, obtains the final classification result, and judges whether the survival sample is an ultraviolet plume target or interference according to the classification result.

[0160] When step 6 is executed, the following steps are specifically included:

[0161] Step 6-1: Call the Transformer encoder output result X trans ;

[0162] Step 6-2: Take the first position output F corresponding to the category embedding:

[0163] F=X trans [:,0];

[0164] Wherein, X trans [:,0] represents the output of the first position;

[0165] is sent to a fully connected layer for classification, and the classification result y is output:

[0166] y=argmax(FC1(F));

[0167] Wherein, FC1 is a single-layer fully connected layer, and argmax is the index of the maximum number;

[0168] y=1 represents an ultraviolet plume target, and y=0 represents interference.

[0169] The embodiment uses the trained Transformer model to distinguish the survival sample to exclude low-altitude complex interference sources. The input data of the Transformer model is the spatiotemporal data intercepted in step 3, that is, the array X containing 90 64x64 pixel matrices. First, an 8-dimensional feature vector is extracted using a convolutional layer, and then position encoding is added, and the Transformer encoder is input, and finally a single-layer fully connected layer is used to obtain the discrimination classification result. The classification result is 1 and 0, respectively, indicating that the survival sample is an ultraviolet plume target and interference.

[0170] The method for detecting the ultraviolet plume target based on the Transformer solves the technical problem of effectively distinguishing the ultraviolet plume target from interference sources such as arc discharge in a low-altitude complex ultraviolet clutter environment, thereby reducing the false alarm rate and improving the detection efficiency. The method uses a two-stage detection architecture. In the first stage, pipeline filtering is used to realize the rapid elimination of noise and most random interference sources by using the target motion continuity and radiation statistical characteristics. In the second stage, a Transformer model based on joint modeling of space-time characteristics is constructed for suspected samples that survive the first stage, which is specially designed to identify low-altitude complex interference sources and effectively reduce the false alarm rate. Instead of directly modeling the entire video, the method uses the precise space-time positioning provided by the pipeline filtering to only intercept the corresponding surviving sample segments as the input of the Transformer, significantly reducing the input data volume. The method uses a single-layer convolution to extract an 8-dimensional spatial gray feature vector to represent each frame of data, taking advantage of the precise positioning of the target. This not only preserves key morphological information but also significantly saves computing resources. The Transformer model combines spatial gray features and time series transformation patterns to accurately depict the differences in dynamic evolution between ultraviolet plume targets and interference sources, thereby improving the discrimination ability. Through the two-stage design of “rough screening first and then precise judgment”, the method achieves a balance between low false alarm rate and real-time performance, making it suitable for engineering applications in low-altitude complex environments.

Claims

1. A Transformer-based method for detecting ultraviolet plume smoke targets, characterized in that: Includes the following steps: Step 1: The data acquisition module acquires a continuous image sequence through the ultraviolet imaging device and sends the image sequence to the pipeline filtering module. The pipeline filtering module performs first-stage pipeline filtering on the image sequence. Based on the motion continuity and statistical radiation characteristics of the ultraviolet plume target, it performs preliminary detection on the suspected target to obtain survival samples, constructs a survival sample set, and outputs it. Step 2: The size regularization module retrieves the survival sample set, retrieves the survival samples from it, and sets a uniform cut-out window based on the spatiotemporal location and size information of the survival samples. The size of the cut-out window is set to C, and the size regularization processing of the generated samples is performed using the cut-out window. Step 3: Within the capture window, the spatiotemporal capture module captures the image region centered on the survival sample from 90 consecutive frames of images, forming a 90-frame pixel matrix array. The spatial size of the pixel matrix array is reduced using bilinear interpolation to construct an array containing 90 64×64 pixel matrices, thus obtaining the spatiotemporal data X of the survival sample. Step 4: The feature extraction module retrieves the spatiotemporal data X and inputs it into the convolutional layer to extract an 8-dimensional spatial feature vector, which represents the spatial grayscale features of the ultraviolet plume target; Step 4 includes the following steps: Step 4-1: Retrieve spatiotemporal data X; Step 4-2: Stitch the spatiotemporal data X horizontally to obtain a large image X. img : X img =hstack(X); Where hstack is the horizontal concatenation operation; Step 4-3: Place X img Remodeled into tensor X tensor : X tensor =reshape(X img ,(1,1,64,5760)); Where, reshape is the reshaping operation, and (1,1,64,5760) are the parameters for reshaping; Step 4-4: For tensor X tensor Perform convolution to extract 8-dimensional feature X conv : X conv =Conv2d(X tensor ,out_channels=8,kernel_size=64,stride=64); Conv2d is a 2D convolution with a kernel size of 64, an output channel count of 8, and a stride of 64. Steps 4-5: Convert the 8-dimensional feature X conv Rearrange as Tensor X = 90 × 8 seq : X seq =rearrange(X conv ,′bchw→b(hw)c′); Where rearrange is the rearrangement operation, 'bchw→b(hw)c' is the parameter of the rearrangement operation, and tensor X seq It contains spatiotemporal features of 90 frames, each frame is represented by an 8-dimensional feature vector; Step 5: The Transformer encoding module performs category embedding on the extracted 8-dimensional feature vectors, then superimposes position encoding, and inputs it into the Transformer encoder for deep spatiotemporal feature modeling; Step 6: The classification and discrimination module extracts the category from the features output by the Transformer encoder and embeds it at the corresponding position. After passing through the fully connected layer, the final classification result is obtained. Based on the classification result, it is determined whether the surviving sample is a UV plume target or interference.

2. The ultraviolet plume target detection method based on Transformer as described in claim 1, characterized in that: During step 1, after receiving a continuous image sequence, the pipeline filtering module processes the image sequence as follows: Step 1-1: Extract the motion trajectory of candidate pixels between consecutive image frames, compare the positional changes of the target in the time dimension, and use the characteristic of continuous motion of ultraviolet plume targets to eliminate isolated noise points in a single image frame. Steps 1-2: Extract ultraviolet radiation intensity within the candidate region, statistically analyze radiation characteristics, and distinguish between noise and suspected targets; Steps 1-3: Based on the motion continuity and radiation characteristics of ultraviolet plume targets, a moving pipe detection method is used to retain only candidate targets that persist within N frames; Steps 1-4: Based on the results of steps 1-3, generate survival samples and record their feature information: Detection time: Detection frame index T; The number of frames present in the moving pipeline: N frames exist; Spatiotemporal location: The center coordinates (x, y) of the suspected target in the frame image t y t ), target size (w) t h t ), where w represents the width, h represents the height, and t is the frame index; Frame index: t = T - N + 1, ..., T - 1, T; Steps 1-5: Based on the results of steps 1-4, generate a set of survival samples.

3. The ultraviolet plume target detection method based on Transformer as described in claim 2, characterized in that: Step 2 includes the following steps: Step 2-1: For each survival sample, take the 90 frames preceding its detection frame T, i.e., t = T−89,…,T, and collect the target's width sequence {w}. t } and height sequence {h t }; If the number of frames N < 90 exists for the survival sample, then the target size in the insufficient frames is set to 0. Step 2-2: Calculate the maximum value of the width sequence: w max =max({w t }); Calculate the maximum value of the altitude sequence: h max =max({h t }); Step 2-3: Determine the side length C of the cropping window: If w max ≤64 and h max If the value is ≤64, then set C=64; otherwise, set C=max(w max ,h max ); Step 2-4: Following the methods in Steps 2-1 to 2-3, generate a square cut-out window with side length C for each survival sample.

4. The ultraviolet plume target detection method based on Transformer as described in claim 3, characterized in that: Step 3 includes the following steps: Step 3-1: For each survival sample, extract the first 89 frames and the detected frames based on the detected frame index T, resulting in a total of 90 consecutive images. The specific frame indices are as follows: t = T−89,…,T−1,T; Step 3-2: In each frame image I t Above, using the center coordinates (x) of the survival samples t y t Using a point as the center, extract a square region with side length C as the extraction window: S t =I t [(x t −C / 2):(x t +C / 2),(y t −C / 2):(y t +C / 2)]; Where St is the spatial data of frame image It; Step 3-3: If the cropping window exceeds the image boundary, the excess part is filled with 0 values. If the number of survival samples is less than 90 frames, the missing frame data is filled with an all-zero matrix. Steps 3-4: Combine the spatial data of consecutive frames to obtain an array S containing 90 C×C pixel matrices: S=[S T−89 ,...,S T−1 ,S T ]; Steps 3-5: Based on the size of the side length C of the captured window, determine whether to perform space reduction processing: If C=64, then take the spatiotemporal data X=S; If C > 64, then for each frame of the image, the matrix S t Perform bilinear interpolation scaling to adjust the size to 64×64, resulting in new spatiotemporal data X: X=[↓S T−89 ,...,↓S T−1 ,↓S T ]; Where ↓ represents the downsampling operation of bilinear interpolation; Steps 3-6: Output spatiotemporal data X.

5. The ultraviolet plume target detection method based on Transformer as described in claim 4, characterized in that: Step 5 includes the following steps: Step 5-1: In tensor X seq Concatenate a category embedding to obtain tensor X cls : X cls =Concat(CLS,X seq ); Where Concat is the concatenation operation and CLS is the learnable category embedding; Step 5-2: Superimpose position information to obtain tensor X pos : X pos =X cls +POS; Where POS stands for location code; Step 5-3: Place X pos The input is fed into a Transformer encoder for deep spatiotemporal feature modeling, yielding the output X of the Transformer encoder. trans : X trans =Transformer(X pos ); Here, Transformer is the Transformer encoder.

6. The method for detecting ultraviolet plume targets based on Transformer as described in claim 5, characterized in that: Step 6 includes the following steps: Step 6-1: Retrieve the Transformer encoder output X trans ; Step 6-2: Output F at the first position corresponding to the category embedding: F=X trans [:,0]; Among them, X trans [:,0] indicates the output of the first position; The data is fed into a fully connected layer for classification, and the classification result y is output: y = argmax(FC1(F)); Where FC1 is a single fully connected layer, and argmax is the index of the maximum number; y=1 represents an ultraviolet plume target, and y=0 represents interference.

Images